MDDI-SCL

MDDI-SCL predicts multi-type drug-drug interactions using supervised contrastive learning to derive and classify drug-pair latent representations for interaction-type prediction.

Key Features:

• Drug feature encoder and mean squared error loss: Employs a self-attention mechanism with an autoencoder to capture drug-level latent features and uses mean squared error (MSE) loss for feature reconstruction.
• Drug latent feature fusion and supervised contrastive loss: Applies multi-scale feature fusion to obtain drug-pair latent features and leverages supervised contrastive learning to cluster similar interaction types and separate dissimilar ones.
• Multi-type DDI prediction and classification loss: Predicts multiple DDI types per drug pair using classification loss for type-specific supervision.
• Evaluation and benchmarking: Evaluated across three tasks on two datasets with performance reported as superior or comparable to state-of-the-art methods.
• Ablation experiments: Uses ablation studies to quantify the contribution of supervised contrastive learning to predictive performance.
• Case studies: Includes case studies demonstrating applicability to real-world drug-drug interaction instances.

Scientific Applications:

• Multi-type DDI identification: Classifies and identifies specific types of drug-drug interactions for polypharmacy risk assessment.
• Mechanistic investigation: Supports elucidation of interaction mechanisms by providing drug-pair latent representations for downstream analysis.
• Pharmacovigilance and patient safety: Aids pharmacovigilance by predicting potential adverse interactions in polypharmacy scenarios.

Methodology:

The method implements three computational modules: a self-attention autoencoder trained with mean squared error loss for drug-level encoding; multi-scale feature fusion with supervised contrastive loss for drug-pair latent representation; and a classifier trained with classification loss for multi-type DDI prediction.